INTERPRETABILITY OF SCANNED AERIAL PHOTOGRAPHS. K.A. Grabmaier, K. Tempfli, R. Ackermann: Scientific Staf, ITC, The Netherlands Girma Messelu: Head Photogrammetry, Ethiopian Mapping Authority, Ethiopia Commission IV, Working Group 3 KEY WORD! Interpretation, Orthoimage, Pixel, Resolution, Feature Extraction, Digital Image, Scanning. ABSTRACT Digital photogrammetry offers a number of advantages, especially towards automation. One of the major sources of digital images for topographic mapping and data collection for spatial databases are aerial photographs. Although interpretability is an important property of images used for this purpose, there is litle known about the loss of interpretability due to scanning of the aerial Photographs. We studied the influence of scanning pixel size on the interpretability of wide angle aerial photographs of scales 1:30,000 and 1:60,000. From the specifications forthe French topographic maps 1:50,000 we selected some eritical point, line- and area-features, Exclusively for these critical features we compared the stereo interpretation of digital images on a Digital Photo- ‘grammetric Workstation (Matra Traster T10) withthe one of the original analogue images on an Analytical Plotter (Zeiss Planicomp C120). Interpretablity of the analogue originals was clearly superior to their digital offsprings with pixel sizes of 60 ym, 30 ym and even 15 4m. Point features suffered the clearest reduction in interpretability To study the influence of pixel size on the usefulness of 1:50,000 orthophoto maps (without annotation), we also produced digital corthophotos from the digital images using different input and output resolutions. We printed them at seale 1:50,000 and tried to interpret from these monoscopic orthophotos the same features asin the stereo interpretation tests, this time with unaided eyes, Point features were not recognizable in ths test series, but for line- and area-features the influence of photoseale and pixel sizes was clear. KURZFASSUNG Die Digitale Photogrammetric bictet cinige Vorteile, insbesondere bezliglich Automatisierung, Der GroBteil der digitalen Bilder, die fir topographische Kartierung und zur Datenerfassung fur raumlichen Datenbanken verwendet werden, sind digitalisiene Lurtbilder. Obwohl die Interpretirbarkeit von Bilder, die fir diesen Zweck verwendet werden, wichtig ist, ist aber den Verlust an Interpretier- barkeit durch das Scannen wenig bekannt, Wir untersuchten den EinfluB der Pixelgrofe auf die Interpretierbarkeit von Weitwinkel- LLuftaufnahmen der Bildmallstibe 1 : 30 000 und 1 : 60 000. Von den Spezifikationen der Franzbsischen topographischen Karte 1: 50 000 haben wir einige kritische Punkt- Linien- und Flichen-Objekte ausgesucht. AusschlieSlich fr diese kritschen Objekte verglichen wir die Stereo-Interpretierbarkeit der digitalen Bilder an cinem Digitalen Auswertegeriit (Matra Traster T10) mit der der analogen Originalbilder an einem Analytischen Auswertegerat (Zeiss Planicomp C120). Die analogen Bilder waren ihren digitalen Derivaten mit Pixelgroflen von 60 um, 30 yum aber auch 15 ym deutch Uberlegen in Interpretierbarkeit. Punktobjekte zeigten dic doutlichste EinbuBe an Interpretierbarkeit Um den EinfluB der PixelgréBe auf die Natzlichkeit von 1 : 50 000 Orthophotokarten (ohne Annotation) zu untersuchen, haben wit von digitalen Bilder verschiedener Pixelgr6en digitale Orthophotos ebenfalls unterschiedlicher PixelgrBen hergestll. Diese wurden im Mafstab 1 : 50 000 ausgegeben, und wir versuchten von den monoskopischen Orthophotos mit unbewaffnetem Auge die gleichen Objekte zu erfassen wie bei der Stereo-Interpretation. Die Punktobjekte waren in dieser Testreihe nicht mu erkennen, aber bei den Linien- und Flachen-Objekten war der Einflu8 von BildmaBstab und Pixelgrben deutlich 2u erkennen. 1, INTRODUCTION (b) the accepted trade off with field completion Digital photogrammetry can offer new approaches to topo- Different from digital satellite images, which offer a fixed spa- ‘graphic data collection and updating for mapping as well as for tal resolution, using aerial photographs allows to influence the spatial databases. A key issue in finding cost-effective solutions, information content of the images through photoscale and scan- however is to assure the required degree of interpretability of ning pixel size. In addition to those Factors which influence the the digital images to be used. This is important for both the image quality ofthe analogue photograph, the scenner properties presently practicable manual and semi-automatic feature and most decisively the pixel size determine the interpretability extraction techniques and the future automated change detection of digital images, which are the input to digital photogrammetric and object recognition processes. solutions to geo-data collection and revision. ‘There is @ whole series of factors influencing interpretability. Several investigations have shown that the digital off-springs can Obviously, the degree of interpretability that is required is compete with their analogue ancestors in geometric accuracy. determined by two aspects: ‘The interpretability aspect, however, has received litte attention (@) the sizes and nature of the objects to be included (this is yet. We may expect lower interpretational quality due to scan strongly influenced by the scale of the map respectively the hing, certainly when using pixel sizes that lead to manageable spatial and thematic resolution of the target database), data volumes on present systems, A reduction in interpretability 305 International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996{implies either a reduction in the information contents of the geo data, or requires an increase of the amount of field completion, ‘The reduction in interpretability could be counteracted by using larger photoscales than we were used to - as long as the implied cost increase ofthe total mapping process does not overtake the expected efficiency gain of using digital instead of analytical photogrammetry. ‘To find proper trade offs for the choosable parameters in a photogrammetric geo-data production line the influence of the ‘mage scale and the scanning pixel size on the interpretability of the digital images should be known, One of the production tines, where digital photogrammetry allows already a far going automation, is image mapping (orthophotography). This gives the possiblity to produce image ‘maps quickly and cheaply, and many developing countries see there a chance to get a complete coverage at scale 1:50,000. Here the interpretability of the ortho images plays an important role for the question how much annotation is required to make the image maps a reasonable substitute for the line maps. Exten- sive annotation however is expensive and time consuming. ‘We carried out a series of interpretational tests to particularly investigate the impact of photoscale and scan resolution, The tests are related to 1:50,000 topographic map specifications and limited to wide-angle photography. We studied interpretability for stereo observation using a digital photogrammetric work- station (Traster T10 of Matra) as well as mono observation with unaided eye of hard copies at scale 1:50,000 of digital ortho- photos (also produced by the T10). 2. MATERIALS AND METHODS 2.1 Images and test site We wanted to use good quality images of two different photo- seales in the usual range for 1:50,000 mapping (1:25,000 to 180,000), preferably from the same site, and the same period, From a site in Souther France we had first generation copies iapositives) of B/W aerial photographs at scales 1:30,000 and 1:60,000 with a difference of only three years and decided to use them for this test. An area of 6km by Tkm was selected, containing varied terrain: flat and billy agricultural parts up t0 rugged mountainous forested parts, It does not contain urban or industrial areas. Itis contained in a single model of the 1:60,000 ‘mages, but from the 1:30,000 photography 4 models are needed to cover it 2.2 Scanning, ‘To make sure, that differences in scan pixel size have a signifi- cant influence on the interpretability, at least the largest pixel size had to give a lower resolution than the original images. To {guarantee this, a "worst case" estimation for the resolution ofthe aerial images was done, With 65% forward overlap and 35% sidelap the maximem radial distance to be used is appr. 110 mm. The camera calibration re- port shows 40 Ip/mm tangential and 49 Ip/mm radial resolution, With aerial film the resolution is probably 20% less than with 308 the film used in the calibration, thus only 32 Ip/mm (tangential). Usually the resolution is determined using high contrast targets (100 : 1), but for interpretability the resolution at low contrast (1.6 = 1), Which may be up to 50% less, is more relevant. To be fn the safe side we used for the "most pessimistic” estimate a value of 16 Ip/mm, thus all used parts of the original diaposi- Lives should have a better resolution than 16 Ip/mm. Scanning with 32 pixels per mm can thus not preserve the resolution full. We could therefor be sure to find a significant difference in resolution between scans with pixel sizes of 60 um (appr. 17 pixels per mm), of 30 um (appr. 33 pixels per mm) and of 15 ym (appr. 67 pixelsmm). ‘Two images (one model) 1:60,000 and six images (four models) 1:30,000 were scanned at GeoRas (Intergraph) with a Zeiss PSI scanner, using a scanning pixel size of 15 ym, We asked Geo- Ras not to cut the tails from the histogram, Pixel sizes of 30 um and 60 ym were oblained by pixel aggregation from the 15 ym mages. Considering the principle of the scanner this can be assumed to be a good simulation of actual scans with 30 ym and 60 ym pixet sizes. 23 Orientation Geometric accuracy was not part of the study. Only orientation errors, which would make it difficult to relate digitized features tnd features in the reference data had to be avoided. The available ground control was far from ideal for our images, but this was no problem. Even large errors, if made consistently in all orientations including the reference data, would have been tolerable, 24 Orthophoto production ‘Automatic DTM generation and ortho-image production of the Matra Traster TIO was used as much as possible with the default parameters. The digital orthophotos were exported to another workstation, enhanced (3x3 Laplace + original image) and then negatives were produced by an Optronics filmwriter and photographic processing. Finally paper prints were made as photographic contact copies. Five types of orthophotos. were produced according to table 1 Input (itl image) Output (orthophoto) ciginal | pixel size on | ortho- | pixel size on scale photo cxiginl | ground | Sate _[orthoph.| ground 1:60,000 | 60m | 36m | 1:50,000 | 100 pm] 5 m 1:60,000 | 30 ym | 18m | 1:50,000 | 100 um | 5 m 160,000 | 30 ym | 1.8m | 1:50.00 | $0 um | 2.5 m 1:60,000| 15 ym | 09m | 1:50,000 | 20 um | 1m 1:30,000] 60 um | 1.8m | 1:50,000 | 50 um | 2.5 m ‘Table 1: Types of Orthophotos From the 1:30,000 photography four separate orthophotos were produced. To avoid any influence of the mosaicing on the test, International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996those pieces were printed separately and individually oriented for the digitizing. For the orthophoto with 20 ym pixel size the negative was produced with 25 um. pixels and photographic re- duction was used instead of contact printing. 25 Data collection Stereo data collections were done by three operators on the Digital Photogrammetric Workstation Traster-T10 from Matra using Demeter as database software. A "coarse to fine" scheme ‘was used such, that remembering the previous interpretation could not bias the new one. When an operator had to interpret ‘an image part again, which he had interpreted before, the new interpretation was with a much higher resolution. scale of | pine size on operator original ‘image | image | ground | wes | cas 160000 | 6oym | 36m | a |B 16900 | 204m | 18m |p | c 13000 | wm | ism | c | a 1:30000 | 30um | o9m | a |p 16000 | 1sum [09m | c | a 13000 | 15um [| oasm [| op | ic 160,000 _| original photographs | another operator 1:30,000 | origina photographs | ca tation training in the neighboring area prior to the actual data collection features: only critical features were selected from the specifica- tions forthe French 1:50,000 map. Features which are probably "always identified” or “never visible" are not considered, Features assumed critical and occurring in the test area were: point features: small isolated houses, bridges, reservoirs, towers, monuments + tine features: foot paths, single lane roads, single track rail ways, ditches, high tension power lines. + area features: vineyards, orchards, cemeteries, All houses up to 120 m? area had to be digitized, To be on the safe side, slightly larger ones had to be digitized too, but not considered in the analysis ‘The same interpretation key was used for the stereo- and for the ‘mono- interpretations. 2 Reference data ‘These were digitized from the original diapositives using the Analytical Plotier Zeiss Planicomp C120 and Kork software for the data collection ‘The same interpretation key was used as for the test data, only hhouses were digitized as area features, and the size limit was bigger. This was necessary to limit the test to houses up 10 120 nv without treating a digitized house which isa litle bigger as an error, In addition to the necessary setting of proper viewing conditions ‘Table 2: Scheme (sequence) for stereo digitizing To allow some comparison between different operators, the area was split in two halves, and each operator did only either the castemn or the western half for a photo seale / pixel size combi nation. Table 2 shows the scheme for the stereo digitizing, in- cluding the collection of the reference data ‘The operators were asked to vary image contrast, brightness and Zoom to be able to interpret the features as good as possible For the mono interpretations there was only a single operator available, and only small differences were expected, so here biases from “memory” could not be excluded. Digitizing tablets With ordinary cursor (no optical magnification) and Microstation software were used. Proper viewing conditions like glare free illumination and a constant viewing distance of 25 em were dif- ficult to achieve, but the major problem for the interpretation was the small area, which was visible around the cursor of the igitizing tablet. The operator stated, that he could clearly ‘identify features on the orthophoto, but often “lost” them when trying to set the cursor there "with a finger on the button”, consistency: all interpretations were done "ignorance based”. Available “other information” like existing topographic maps or higher resolution images were not to be used. The operators ‘were familiar with this terrain type, they had done map comple tion exercises not far from the site at some time before. To achieve a reasonable consistency they carried out an interpre- (eye base, focussing, image rotation and squint) operators were asked to and to vary the illumination the magnification to get the interpretations as good as possible 27 Analysis Method Data digitized from the orthophotos or on the Digital Photo- grammetric Workstation Matra Traster T10 were compared with igitizations from the original diapositives on the Analytical Plotter Zeiss Planicomp C120. Thus the reference data for the ‘two image scales had to be different. This was necessary, be- ‘cause there were obviously significant changes in the three years between the flights All data were transferred to Arcinfo, creating separate coverages for each feature in each digitization. After cleaning of the data (eg. all houses > 120 m had to be identified and eliminated from the analysis) matches between the test data and the refe- rence data were identified and counted for point features or the lengths added up for line features respectively. For area features intersections of the polygons were used For different sets of test data (6 stereo digitizations from 2 image scales, each with 3 pixel sizes - with and without sub- division in two parts for different operators - and 5 orthophoto types) separate "confusion matrices” for point features, line fea- tures and area features respectively were constructed. In the ‘mono interpretations point features could not be identified, so these confusion matrices were not set up, Table 3 gives. an example of a confusion matrix for 4 features (1 t0 4), 307 International Archives of Photogrammetry and Remote Sensing. Vol. XXX, Part B4. Vienna 1996